AI forecasting experiments show mixed results against human forecasters
Reported by Tech Times1 source ↓
AI agents and commercial models have made forecasts in fantasy soccer and prediction markets, but results remain inconsistent. One market-wagering experiment lost money, and skilled human forecasters remain difficult to outperform.
Key points
- A WIRED fantasy soccer test improved after agents were assigned specialized research and analysis roles
- Adding more than 100 agents did not significantly improve forecasts and mainly raised API costs
- In one comparison, GPT-5 ranked ahead of Grok 4, Claude Sonnet, Gemini and Llama
- Models given $10,000 to wager over 57 days lost money, despite signs of improving performance
Details
Tech Times reports that WIRED tested whether a group of AI agents could build and manage a competitive Premier League fantasy football team. Early recommendations missed player and team considerations, including leaving Erling Haaland on the bench. Results improved after the system added agents for news gathering, performance analysis, modeling and reviewing recommendations.
The report says increasing the system to more than 100 agents did not significantly improve results and mainly increased API costs.
A separate Prediction Arena experiment involving Harvard researchers and Arcada Labs gave frontier models $10,000 to wager in prediction markets over 57 days. The models lost money, though researchers saw signs of improving performance. The article distinguishes estimating probabilities from consistently making profitable trades, and says skilled human forecasters remain difficult to outperform.
Written from coverage as of
Images
- Tech Times